Industries / Industrial equipment and services / Industrial services / Commercial diving
Inspection
Underwater inspection is the work of cleaning, looking at and measuring the submerged parts of platforms, pipelines, ship hulls, bridges, dams and intakes, and recording what is found.
6
Milestones achieved, from 1832 to 2025.
1
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1832
A bridge surveyed by diver
In December 1832 the London civil engineers Walker and Burges, asked to report on the soundness of the ageing Blackfriars Bridge across the Thames, hired Charles Deane to survey its underwater parts in his new air-fed diving helmet, and tried the equipment themselves. John Bevan, a physiologist who wrote a history of the Deane brothers, calls it the first use of the diving helmet in civil engineering.
Credit
Charles Anthony Deane; the engineers Walker and Burges; account by John Bevan
1988
Bridge footings checked under water
A 1981 report for the US National Cooperative Highway Research Program found that only 15 states routinely inspected their bridges below the waterline. On 5 April 1987 the Schoharie Creek Bridge on the New York State Thruway fell into a flooded creek after the water scoured away the soil under a pier, and 10 people died. The bridge had been inspected every one or two years since 1968, but its pier footings had never been examined under water; an underwater inspection was scheduled for 1987. Revised National Bridge Inspection Standards, published on 26 August 1988, made states list every bridge that needs underwater inspection and record it in the national inventory, and the Federal Highway Administration's guidance says the revision requires those inspections at least every five years, by divers where the water is too deep to wade.
LessonA bridge can pass every inspection above water while its foundation washes out.
Credit
US Federal Highway Administration; National Transportation Safety Board (investigation); Wiss, Janney, Elstner Associates (investigation for the Thruway Authority)
1995
Weld checks move to ROVs
Divers checked platform welds for cracks with magnetic particle inspection (MPI), in which iron powder gathers over a crack in magnetized steel; it needs the steel cleaned to bare metal. G. A. Raine of Technical Software Consultants wrote in 1995 that ROVs (remotely operated vehicles, uncrewed submarines piloted through a cable) built in the 1980s could already clean, look and run flooded member detection, a test for water inside a hollow brace that shows a crack has gone right through its wall, and that these jobs had cut diving to about 10% of combined diving and ROV programmes. Welds were the job left. His firm then built an array head for alternating current field measurement (ACFM, a crack-sizing method described under Non-destructive testing) for a North Sea operator; any work-class ROV could carry it without tracking the weld precisely, and it passed tank trials of detection and sizing.
Credit
Technical Software Consultants (G. A. Raine)
2000s
Fewer looks, fewer finds
Norway's petroleum regulator keeps CODAM, a public database of corrosion and damage reported on fixed and floating installations on the Norwegian shelf since the mid-1970s. Analysing its 4,639 reports from 1974 to 2021 for the Petroleum Safety Authority, the engineering firm Ramboll found that 58.4% had been found by magnetic particle inspection and 30.1% by visual methods, ROV inspection among them. Reports rose from about 1980 and fell sharply from about 2000, when operators moved from annual inspections to plans set by risk and reliability; Ramboll gives fewer inspections and safer structures as the two possible reasons.
LessonWhen inspection plans thin out, a falling count of findings cannot by itself show that structures got safer.
Credit
Petroleum Safety Authority Norway (database); Bjørn T. Svendsen, Ramboll (analysis, 2022)
2021
Hull surveyed by mini ROV
In August 2021 Bureau Veritas (BV), a classification society (a firm that sets and checks ships' technical standards), and the drone company MaDfly completed what they called the first full in-water survey of a ship's hull bottom with a mini ROV, on Brittany Ferries' Bretagne, supervised by BV for the French flag. BV ran the survey a second time with live video and no surveyor on board.
Credit
Bureau Veritas; MaDfly – Marine Drone Services
2025
ROVs named in hull surveys
On 30 May 2025 the US Coast Guard issued NVIC 01-89 Change 1, the first major revision of its underwater survey guidance for US-flagged vessels since 1989. It explicitly allows ROVs to carry out underwater hull surveys, and lets classed vessels under 15 years old do them without separate Coast Guard approval.
Credit
US Coast Guard, Office of Commercial Vessel Compliance
Watch for
An AUV inspects a jacket
BeeX says its A.IKANBILIS, an autonomous underwater vehicle (AUV, a robot submarine that runs without a pilot), inspected a four-legged oil and gas jacket (the steel lattice that holds a platform up) in 71 m of water for a major operator in June 2026. It reports a general visual inspection, ten cathodic protection readings (voltage checks that show whether the anodes still protect the steel), about 300 anodes, 150 braces, three J-tubes, risers and clamps, in 4.5 operating days on one shift.
Achieved whenthe operator or a certifying body accepts the results as the jacket's required underwater inspection, or a comparison with an ROV or diver survey of the same structure is published.
Claimed by
BeeX, June 2026, as summarized by ShipUniverse
Open
01
Overdue dives, bridge by bridge
A public table, built from the Federal Highway Administration's National Bridge Inventory (NBI) files, of every US bridge that needs an underwater inspection: its required interval, the date of its last underwater inspection, its scour-critical rating, and a flag where the last inspection is older than the interval allows.
The briefCredit · unclaimed02
Test set for survey logs
A public set of pipeline survey video frames labeled with the events that survey staff log, such as burial, exposure, field joints, anodes and free spans (stretches where the pipe hangs above the seabed), and a score for AI models on it, including general vision-language models used without extra training.
The briefCredit · unclaimed
Open 01 · the brief
Overdue dives, bridge by bridge
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public table, built from the Federal Highway Administration's National Bridge Inventory (NBI) files, of every US bridge that needs an underwater inspection: its required interval, the date of its last underwater inspection, its scour-critical rating, and a flag where the last inspection is older than the interval allows.
Why it is still open
Federal rules set underwater intervals of at most 24, 60 or 72 months by risk, and the NBI files carry each bridge's underwater interval (item 92B) and last underwater inspection date (item 93B). FHWA scores each state on underwater intervals (its metrics 8 and 9) in annual reviews, but searches found no published bridge-by-bridge list of late underwater inspections and no study that computes one.
Built with what exists
FHWA's delimited NBI files to 2025 (fields UNDWATER_LOOK_SEE_092B, UNDWATER_LAST_DATE_093B and SCOUR_CRITICAL_113); Python with pandas or DuckDB; an AI coding agent to write and test the date logic; public records requests to check flagged bridges with their owners.
First step
Load the 2024 Delaware file (872 records, 128 of them coded as needing underwater inspection), compute each one's due date, and list those past due.
Done looks like
Every state, the District of Columbia and Puerto Rico for the 2018 to 2025 files; the count and share of bridges needing underwater inspection whose last one was older than their interval when the file was submitted; 50 flagged bridges checked with their owners and the share confirmed late published.
Open 02 · the brief
Test set for survey logs
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public set of pipeline survey video frames labeled with the events that survey staff log, such as burial, exposure, field joints, anodes and free spans (stretches where the pipe hangs above the seabed), and a score for AI models on it, including general vision-language models used without extra training.
Why it is still open
On a survey vessel a data coordinator calls out events by voice as the ROV flies the pipeline, and staff onshore check the log, which the authors of a 2020 study call slow, costly and prone to human error. That study, by Anastasios Stamoulakatos and colleagues at the University of Strathclyde with the survey firm N-Sea, trained a network on 23,570 frames from two North Sea surveys and reported 95.1% to 99.7% accuracy per frame, but on the firm's own footage, which it did not release. Olavo Sampaio's 2020 master's thesis at the Federal University of Rio de Janeiro labeled 457,000 frames from 146 pipeline inspection videos for anodes, burial, damage, flanges and repairs, and did not release them either, because the source videos are proprietary. SubPipe (2024), recorded by an AUV over a sea outfall, is public but labels image regions and sonar objects, not survey events, and annotated ROV footage found elsewhere is sold commercially.
Built with what exists
SubPipe's open camera footage; survey video released by contractors or operators, such as from decommissioned pipelines; CVAT for labeling; a vision-language model such as Claude Fable 5.2 (see the Note "Three working devices") or Gemini, and Gemini Decision 1 to answer fixed questions about each frame (see the Note "A thousand questions, one pass"); the Strathclyde ResNet-50 recipe as a baseline.
First step
Ask five survey contractors or operators to release one survey of a retired pipeline under an open license, and meanwhile label 1,000 SubPipe frames for exposure and burial.
Done looks like
20,000 frames from at least three surveys by different contractors, labeled by two trained people with their agreement published, and a per-event F1 score (a measure that balances missed and false events) for at least three methods on footage none of them was trained on.